Meta says it won’t sign Europe AI agreement, calling it an overreach

One of the key aspects of the act is how a model provider is responsible if the downstream partners misuse it in any way. For open source, it's a very hard requirement[1].

> GPAI model providers need to establish reasonable copyright measures to mitigate the risk that a downstream system or application into which a model is integrated generates copyright-infringing outputs, including through avoiding overfitting of their GPAI model. Where a GPAI model is provided to another entity, providers are encouraged to make the conclusion or validity of the contractual provision of the model dependent upon a promise of that entity to take appropriate measures to avoid the repeated generation of output that is identical or recognisably similar to protected works.

[1] https://www.lw.com/en/insights/2024/11/european-commission-r...

This is European law, not US. Reasonable means reasonable and judges here are expected to weigh each side's interests and come to a conclusion. Not just a literal interpretation of the law.

Same way the pols aren’t futurists and perfect neither is anyone else. Everyone should sit at the table and discuss this like adults.

You want to go live in the hills alone, go for it, Dick Proenneke. Society is people working collectively.

Blaming tools for the actions of their users is stupid.

LLMs are hardly reliable ways to reproduce copyrighted works. The closest examples usually involve prompting the LLM with a significant portion of the copyrighted work and then seeing it can predict a number of tokens that follow. It’s a big stretch to say that they’re reliably reproducing copyrighted works any more than, say, a Google search producing a short excerpt of a document in the search results or a blog writer quoting a section of a book.

It’s also interesting to see the sudden anti-LLM takes that twist themselves into arguing against tools or platforms that might reproduce some copyrighted content. By this argument, should BitTorrent also be banned? If someone posts a section of copyrighted content to Hacker News as a comment, should YCombinator be held responsible?

In some cases they can be prompted to guess a number of tokens that follow an excerpt from another work.

They do not contain all copyrighted works, though. That’s an incorrect understanding.

Commercial use of someone's image also already has laws concerning that as far as I know, don't they?

https://en.wikipedia.org/wiki/Printer_tracking_dots

Only because the companies are intentionally making it so. If they weren't trained to not reproduce copyrighted works they would be able to.

They're not really "blaming" the tool though. They're using a supply chain attack against the subset of users they're interested in.

For reference, see every highly-regulated industry everywhere.

You think Sam Altman was in testifying to the US Congress begging for AI regulation because he's just a super nice guy?

In the end the regulation was diluted to something that made sense(ish) but that process took about 4 years. It also slowed down all enterprise deals because no one knew if a deal was going to be against GDPR and the lawyers defaulted to “no” in those orgs.

Asking regulators to understand and shape market evolution in AI is basically asking them to trade stocks by reading company reports written in mandarin.

So I disagree, probably Europe will be even more behind in ev if they doesn't push eu manufacturers to invest so heavily in the industry.

You can se for example than for legacy manufacturers the only ones in the top ten are Europeans being 3 out of 10 companies, not Japanese or Korean for example, and in Europe Volkswagen already overtake Tesla in sales Q1 for example and Audi isn't that much away also.

The YouTube deal was a lot earlier than Instagram, 2006. Google was way smaller than now. iPhone wasn’t announced. And it wasn’t two social networks merging.

Very hard to see how regulators could have the clairvoyance to see into this specific future and its counter-factual.

In those places actually fees are included ("reprographic levy") in the appliance, and the needed supply prices, or public operators may need to pay additionally based on usage. That money goes towards funds created to compensate copyright holders for loss of profit due to copyright infringement carries out through the use of photocopiers.

Xerox is in no way singled out and discriminated against. (Yes, I know this is an Americanism)

Doomers have been wrong about completely different doom scenarios in the past (+), but it says nothing about to this new scenario. If you're doing statistics in your head about it, you're wrong. We can't use scenarios from the past to make predictions about completely novel scenarios like thinking computers.

(+) although they were very close to being right about nuclear doom, and may well be right about climate change doom.

If you are paying for lobbyists to write the legislation you want, as corporations do, you get the law you want - that excludes competition, funds your errors etc.

The point is you are not dealing with 'humanity', you are dealing with those who represent authority for humanity - not the same thing at all. Connected politicians/CEOs etc are not actually representing 'humanity' - they merely say that they are doing so, while representing themselves.

Is the same regulation that was introduced in 2016. The only people who pretend not to understand it are those who think that selling user data to 2000+ "partners" is privacy

Innovation is good, but there's no need to go as fast as possible. We can be careful about things and study the effects more deeply before unleashing life changing technologies into the world. Now we're seeing the internet get destroyed by LLMs because a few people decided it was ok to do so. The benefits of this are not even clear yet, but we're still doing it just because we can. It's like driving a car at full speed into a corner just to see what's behind it.

AFAICT the actual text of the act[0] does not mention anything like that. The closest to what you describe is part of the chapter on copyright of the Code of Practice[1], however the code does not add any new requirements to the act (it is not even part of the act itself). What it does is to present a way (which does not mean it is the only one) to comply with the act's requirements (as a relevant example, the act requires to respect machine-readable opt-out mechanisms when training but doesn't specify which ones, but the code of practice explicitly mentions respecting robots.txt during web scraping).

The part about copyright outputs in the code is actually (measure 1.4):

> (1) In order to mitigate the risk that a downstream AI system, into which a general-purpose AI model is integrated, generates output that may infringe rights in works or other subject matter protected by Union law on copyright or related rights, Signatories commit:

> a) to implement appropriate and proportionate technical safeguards to prevent their models from generating outputs that reproduce training content protected by Union law on copyright and related rights in an infringing manner, and

> b) to prohibit copyright-infringing uses of a model in their acceptable use policy, terms and conditions, or other equivalent documents, or in case of general-purpose AI models released under free and open source licenses to alert users to the prohibition of copyright infringing uses of the model in the documentation accompanying the model without prejudice to the free and open source nature of the license.

> (2) This Measure applies irrespective of whether a Signatory vertically integrates the model into its own AI system(s) or whether the model is provided to another entity based on contractual relations.

Keep in mind that "Signatories" here is whoever signed the Code of Practice: obviously if i make my own AI model and do not sign that code of practice myself (but i still follow the act requirements), someone picking up my AI model and signing the Code of Practice themselves doesn't obligate me to follow it too. That'd be like someone releasing a plugin for Photoshop under the GPL and then demanding Adobe release Photoshop's source code.

As for open source models, the "(1b)" above is quite clear (for open source models that want to use this code of practice - which they do not have to!) that all they have to do is to mention in their documentation that their users should not generate copyright infringing content with them.

In fact the act has a lot of exceptions for open-source models. AFAIK Meta's beef with the act is that the EU AI office (or whatever it is called, i do not remember) does not recognize Meta's AI as open source, so they do not get to benefit from those exceptions, though i'm not sure about the details here.

[0] https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:...

[1] https://ec.europa.eu/newsroom/dae/redirection/document/11811...

Preferable to a burgeoning oligarchy.

The other thing is that approximately all of the training set is copyrighted, because that's the default even for e.g. comments on forums like this comment you're reading now.

The other other thing is that at least two of the big model makers went and pirated book archives on top of crawling the web.

Personally, I don’t think they’re bad. Plastic isn’t that harmful, and neither is social media.

I think people romanticize the past and status quo. Change is scary, so when things change and the world is bad, it is easy to point at anything that changed and say “see, the change is what did it!”

It's not just IT. Ask any EU farmer.

It was made impractical by ad platforms and others who decided to use dark patterns, FUD and malicious compliance to deceive users into agreeing to be tracked.

Instead of exactly as you say: a global browser option.

As someone who has had to implement this crap repeatedly - I can’t even begin to imagine the amount of global time that has been wasted implementing this by everyone, fixing mistakes related to it and more importantly by users having to interact with it.

The law is written to encourage such defaults if anything, it just wasn't profitable enough I guess.

This is handwaving. We can be pretty well sure at this point what the effects aren’t, given their widespread prevalence for generations. We have a 2+ billion sample size.

Im fully supportive of consent, but the way it is implemented is impractical from everyone’s POV and I stand by that.

But with a little bit of reading, one could ultimately summarise the enormous wall of text simply as: “We’ve added your email address to a marketing list, click here to opt out.”

The huge wall of text email was designed to confuse and obfuscate as much as possible with them still being able to claim they weren’t breaking protection of personal information laws.

[0]: https://imgur.com/a/aN4wiVp

Technically untrue, monopoly busting is a kind of regulation. I wouldn't bet on it happening on any meaningful scale, given how strongly IT benefits from economies of scale, but we could be surprised.

I don't think practical is the right word here. All the businesses in the world operated without tracking until the mid 90s.

Do you think Heroin is good?

It is pretty clear

If you cannot see the difference between BitTorrent and Ai models, then it's probably not worth engaging with you.

But Ai model have been shown to reproduce the training data

https://gizmodo.com/ai-art-generators-ai-copyright-stable-di...

https://arxiv.org/abs/2301.13188

But yes, perhaps they should have worked with e.g. Mozilla to develop some kind of standard browser interface for this.

I'm constantly clicking away cookie banners on UK government or NHS (our public healthcare system) websites. The ICO (UK privacy watchdog) requires cookie consent. The EU Data Protection Supervisor wants cookie consent. Almost everyone does.

And you know why that is? It's not because they are scammy ad funded sites or because of government surveillance. It's because the "cookie law" requires consent even for completely reasonable forms of traffic analysis with the sole purpose of improving the site for its visitors.

This is impractical, unreasonable, counterproductive and unintelligent.

The conversation is not about my opinion on tracking, anyway. It’s about the impracticality of implementing the legislation that is hostile and time consuming for both website owners and users alike

To be clear, I don't have any particular insight on whether this is possible right now with LLMs, and I'm not taking a stance on copyright law in general with this comment. I don't think your argument makes sense though because there's a clear technical difference that seems like it would be pretty significant as a matter of law. There are plenty of reasonable arguments against things like the agreement mentioned in the article, but in my opinion, your objection isn't one of the.

You just made the model not open source

In a rigidly regulated market with preemptive action by regulators (like EU, Japan) you end up with a persistent oligarchy that is never replaced. An aristocracy of sorts.

The middle road is the best. Set up a fair playing field and rules of the game, but allow innovation to happen unhindered, until the dust has settled. There should be regulation, but the rules must be bought with blood. The risk of premature regulation is worse.

While some people get addicted to it, the vast majority of users are not addicts. They choose to use it.

It keeps the political grifters who make these regulations employed, that's kind of the main point in EU/UKs endless stream of regulations upon regulations.

WHAT?! Do you think we as humanity would have gotten to all the modern inventions we have today like the internet, space travel, atomic energy, if we had skipped the fossil fuel era by preemptively regulating it?

How do you imagine that? Unless you invent a time machine, go to the past, and give inventors schematics of modern tech achievable without fossil fuels.

The same is true of privacy policies. I’ve seen some companies have very short policies I could read in less than 30s, those companies are not suspicious.

Yup. That's what those 2000+ "partners" are all about if you believe their "legitimate interest" claims: "improve traffic"

Oh, we already know large chunks of it, and the regulations explicitly address that.

If the chest-beating crowd would be presented with these regulations piecemeal, without ever mentioning EU, they'd probably be in overwhelming support of each part.

But since they don't care to read anything and have an instinctive aversion to all things regulatory and most things EU, we get the boos and the jeers

This is a personal decision to be made by the data "donor".

The NHS website cookie banner (which does have a correct implementation in that the "no consent" button is of equal prominence to the "mi data es su data" button) says:

> We'd also like to use analytics cookies. These collect feedback and send information about how our site is used to services called Adobe Analytics, Adobe Target, Qualtrics Feedback and Google Analytics. We use this information to improve our site.

In my opinion, it is not, as described, "completely reasonable" to consider such data hand-off to third parties as implicitly consented to. I may trust the NHS but I may not trust their partners.

If the data collected is strictly required for the delivery of the service and is used only for that purpose and destroyed when the purpose is fulfilled (say, login session management), you don't need a banner.

The NHS website is in a slightly tricky position, because I genuinely think they will be trying to use the data for site and service improvement, at least for now, and they hopefully have done their homework to make sure Adobe, say, are also not misusing the data. Do I think the same from, say, the Daily Mail website? Absolutely not, they'll be selling every scrap of data before the TCP connection even closes to anyone paying. Now, I may know the Daily Mail is a wretched hive of villainy and can just not go there, but I do not know about every website I visit. Sadly the scumbags are why no-one gets nice things.

My problem is that users cannot make this personal decision based on the cookie consent banners because all sites have to request this consent even if they do exactly what they should be doing in their users' interest. There's no useful signal in this noise.

The worst data harvesters look exactly the same as a site that does basic traffic analysis for basic usability purposes.

The law makes it easy for the worst offenders to hide behind everyone else. That's why I'm calling it counterproductive.

[Edit] Wrt NHS specifically - this is a case in point. They use some tools to analyse traffic in order to improve their website. If they honour their own privacy policy, they will have configured those tools accordingly.

I understand that this can still be criticised from various angles. But is this criticism worth destroying the effectiveness of the law and burying far more important distinctions?

The law makes the NHS and Daily Mail look exactly the same to users as far as privacy and data protection is concered. This is completely misleading, don't you think?

I think you've got civil and common law the wrong way round :). US judges have _much_ more power to interpret law!

This part gave me a genuine laugh. Good joke.

Drug trafficking, stealing, scams are massive revenue for gangs.

adjusts tinfoil hat

That's an awfully callous approach, and displays a disturbing lack of empathy toward other people.

> When interpreting EU law, the CJEU pays particular attention to the aim and purpose of EU law (teleological interpretation), rather than focusing exclusively on the wording of the provisions (linguistic interpretation).

> This is explained by numerous factors, in particular the open-ended and policy-oriented rules of the EU Treaties, as well as by EU legal multilingualism.

> Under the latter principle, all EU law is equally authentic in all language versions. Hence, the Court cannot rely on the wording of a single version, as a national court can, in order to give an interpretation of the legal provision under consideration. Therefore, in order to decode the meaning of a legal rule, the Court analyses it especially in the light of its purpose (teleological interpretation) as well as its context (systemic interpretation).

https://www.europarl.europa.eu/RegData/etudes/BRIE/2017/5993...

If they only do this, they don't need to show anything.

Meta, OpenAI, Nvidia, Microsoft and Google don't care about people. They care about control: controlling influence, knowledge and universal income. That's the endgame.

Just like in the US, the EU has brilliant people working on regulations. The difference is, they're not always working for the same interests.

The world is asking for US big tech companies to be regulated more now than ever.

Space travel does need a lot of oil, so it might be affected, but the beginning of it were in the 40s so the research idea was already there.

Atomic energy is also from the 40s and might have been the alternative to oil, so it would thrive more if we haven't used oil that much.

Also all 3 ARE heavily regulated and mostly done by nation states.

I'm not sure why you and GP are trying to use this point to draw a contrast to the US? That very much is a feature in US law as well.

And this is the crux of the problem. The law helps a tiny minority of people enforce an extremely (and in my view pointlessly) strict version of privacy at the cost of misleading everybody else into thinking that using analytics for the purpose of making usability improvements is basically the same thing as sending personal data to 500 data brokers to make money off of it.

You seem to not have understood that different forms of appliances need to comply with different forms of law. And you being able to call it open source or not doesn't change anything about its legal aspects.

And every law written is a compromise between two opposing parties.

What exactly do think should be allowed which still respect privacy, which isn't now?

What you could then add to this system is a certification scheme to permit implicit consent of all the data handling (including who you hand it off to and what they are allowed to do with it, as well as whether they have demonstrated themselves to be trustworthy) is audited to be compliant with some more stringent requirements. It could even be self-certification along the lines of CE marking. But that requires strict enforcement, and the national regulators so far have been a bunch of wet blankets.

That actually would encourage organisations to find ways to get the information they want without violating the privacy of their users and anyone else who strays into their digital properties.

Anyway, the show must go on so were unlikely to see any reversal of this. It’s a big experiment and not necessarily anything that will benefit even the model providers themselves in the medium term. It’s clear that the ”free for all” policy on grabbing whatever data you can get is already having chilling effects. From artists and authors not publishing their works publicly, to locking down of open web with anti-scraping. Were basically entering an era of adversarial data management, with incentives to exploit others for data while protecting the data you have from others accessing it.

Technology changes very quickly and the future of things is hardly decided by entrenched interests.

There's a certain hubris to applying rules and regulations to a system that you fundamentally don't understand.

You're hypothesizing the existence of large negative effects with minimal evidence.

But the positive effects of plastics and social media are extremely well understood and documented. Plastics have revolutionized practically every industry we have.

With that kind of pattern of evidence, I think it makes sense to discount the negatives and be sure to account for all the positives before saying that deploying the technology was a bad idea.

In limited, closed systems, it may not escape, but all it takes is one bad (or hacked) actor and the privacy of it is gone.

In a way, we used to be "protected" because it was "too big" to process, store, or access "everything".

Now, especially with an economic incentive to vacuum literally all digital information, and many works being "digital first" (even a word processor vs a typewriter, or a PDF that is sent to a printer instead of lithograph metal plates)... is this the information Armageddon?

Simply taking away some giant precursor for the advancements we enjoy today and then assuming it all would have worked out somehow is a bit naive.

I would need to see a very detailed pipeline from growing wheat in an agrarian society to the development of a microprocessor without fossil fuels to understand the point you're making. The mining, the transport, the manufacture, the packaging, the incredible number of supply chains, and the ability to give people time to spend on jobs like that rather than trying to grow their own food are all major barriers I see to the scenario you're suggesting.

The whole other aspect of this discussion that I think is not being explored is that technology is fundamentally competitive, and so it's very difficult to control the rate at which technology advances because we do not have a global government (and if we did have a global government, we'd have even more problems than we do now). As a comment I read yesterday said, technology concentrates gains towards those who can deploy it. And so there's going to be competition to deploy new technologies. Country-level regulation that tries to prevent this locally is only going to lead to other countries gaining the lead.

LLMs extract semantic information from their training data and store it at extremely low precision in latent space. To the extent original works can be recovered from them, those works were nothing intrinsically special to begin with. At best such works simply milk our existing culture by recapitulating ancient archetypes, a la Harry Potter or Star Wars.

If the copyright cartels choose to fight AI, the copyright cartels will and must lose. This isn't Napster Part 2: Electric Boogaloo. There is too much at stake this time.

Your point is fundamentally philosophical, which is you can't use the past to predict the future. But that's actually a fairly reductive point in this context.

GP's point is that simply making an argument about why everything will fail is not sufficient to have it be true. So we need to see something significantly more compelling than a bunch of arguments about why it's going to be really bad to really believe it, since we always get arguments about why things are really, really bad.

They've just mastered the art of lying to gullible idiots or complicit psycophants.

It's not new to anyone who pays and kind of attention.

It's not like users are accidentally producing copies of Harry Potter.

But it is also increasingly dubious that the public gets a good deal out of copyright law anyway.

> From artists and authors not publishing their works publicly

The vast majority of creators have never been able to get remotely close to make a living from their creative work, and instead often when factoring in time lose money hand over fist trying to get their works noticed.

Historically, all nations that developed and deployed new tech, new sources of energy and new weapons, have gained economic and military superiority over nations who did not, which ended up being conquered/enslaved.

UK would not have managed to be the world power before the US, without their coal fueled industrial era.

So as history goes, if you refuse to take part in, or cannot keep up in the international tech, energy and weapons race, you'll be subjugated by those who win that race. That's why the US lifted all brakes on AI, to make sure they'll win and not China. What EU is doing, self regulating itself to death, is ensuring its future will be at the mercy of US and China. I'm not the one saying this, history proves it.

> > GPAI model providers need to establish reasonable copyright measures to mitigate the risk that a downstream system or application into which a model is integrated generates copyright-infringing outputs, including through avoiding overfitting of their GPAI model. Where a GPAI model is provided to another entity, providers are encouraged to make the conclusion or validity of the contractual provision of the model dependent upon a promise of that entity to take appropriate measures to avoid the repeated generation of output that is identical or recognisably similar to protected works.

It sounds to me like an LLM you describe would be covered if they people distributing it put in a clause in the license saying that people can't do that.

Your augment only work in a fictional world where oil does not exist and you have the hindsight of today.

But when oil does exist and if you would have chosen not to use it, you will have long been steamrolled by industrialized nations powers who used their superior oil fueled economy and military to destroy or enslave your nation and you wouldn't be writing this today.

> How would you have won the world wars without oil?

You don't need to win world wars to have technological advancement, in fact my country didn't. I think the problem with this discussion, is that we all disagree what to regulate, that's how we ended up with the current situation after all.

I interpreted it to mean that we wouldn't use plastic for everything. I think we would be fine having glass bottles and paper, carton, wood for grocery wrapping. It wouldn't be so individual per company, but this not important for the economy and consumers, and also would result in a more competitive market.

I also interpreted it to mean that we wouldn't have so much cars and don't use planes beside really important stuff (i.e. international politics). The cities simply expand to the travel speed of the primary means of transportation. We would simply have more walkable cities and would use more trains. Amazon probably wouldn't be possible and we would have more local producers. In fact this is what we currently aim for and it is hard, because transition means that we have larger cities then we can support with the primary means of transportation.

As for your example inventions: we did have computers in the 40s and the need for networking would arise. Space travel is in danger, but you can use oil for space travel without using it for everyday consumer products. As I already wrote, we would have more atomic energy, not sure if that would be good though.

The moment the EU shows even a small sign of protectionism, the US complains. It's a double standard.

Why take such risks when we could take our time doing more studies and thinking about all the possible scenarios? If we did, we might use plastics where they save lives and not use them in single-use containers and fabrics. We'd get most of the benefit without any of the harm.

But if such co-operation was possible, it would make sense to progress more carefully.